Chapter 6

Economic Selection Indexes

In economically oriented breeding programs, the trait that we want to improve could be called economic merit. The breeding objective of our program is then to maximize improvement of economic merit. Economic merit might be defined in different ways, e.g. as profit per animal or profit per enterprise or economic efficiency or something else. We will return to this problem in later Chapters. For the present, it is only necessary to recognize that the breeding objective is a general statement of the economic genetic goal of the breeding program. For a given definition of the breeding objective, there will likely be several or many traits, which would contribute to the objective. The aggregate genotype is then defined as a function of the additive genetic values of the traits of interest of an individual, which if selected upon would achieve the breeding objective. The function need not necessarily be linear, but in many cases an approximate linear relationship can be found which adequately defines aggregate genotype over the range of genetic values encountered (see Chapter 7). If the function is a linear function, then the aggregate genotype, H, can be written as H = v1g1 + v2g2 vngn = vg (6.1)

where gi is the additive genetic value of the ith trait and vi is the economic value of genetically improving that trait. Note that vi is a partial economic weight, that is, it is the economic value of genetically improving the ith trait, when all other traits remain unchanged. In vector notation, v = [ v1 , v2 , , vn] and g = [ g1 , g2 , , gn] . In practice, the additive genetic value of the various traits for each individual are not known. However we can record each individual's performance for a number of traits. The observations on these traits can then be combined into a selection index, I of the form, I = b1x1 + b2x2 bmxm = bx (6.2)

where xi is an observation on the ith trait and bi is the selection index coefficient (or weight) for that trait. In vector notation, b = [ b1 , b2 , , bm] and x = [ x1 , x2 , , xm]. The problem is then to estimate the selection index coefficients, bi, such that selection of individuals on their selection index value, I, maximizes response in the aggregate genotype, H. The selection index methods described in Chapter 4 can be used to derive such indexes based on: b = P-1Gv (6.3)

6-1

General Derivation of Optimal Weights

Selection Index Theory

Economic Breeding Objective/Goal

H = v1 g1 + v2 g2 + . . + vn gn = vg
gi = true bv for trait i vi = economic value trait i H=g for single trait EBV

Selection Index

I = b1 x1 + b2 x2+ . . + bm xm= bx
xi = phenotypic record bi = index weight Derive weights to maximize - response in H: RH=bH,I(I - I) - accuracy of I: rI,H

6-2

It is worth noting at this point that the traits recorded and which appear in the index do not need to be, and often are not, the same traits as appear in the aggregate genotype. As a crude example, consider a terminal sire line of pigs. It might be found that the profitability of this line was (approximately) a linear function of carcass weight and lean percentage at 120 days of age. Thus the aggregate genotype would be written as: H = v1 (120 day carcass weight) + v2 (120 day lean percentage) Neither carcass weight nor lean percentage can be recorded directly in live pigs, but we could record live weight at 120 days and ultrasonically estimated back fat depth at 120 days, and the selection index would take the form, I = b1 (120 day live weight) + b2 (120 day ultrasonic back fat depth) In this case, neither of the two traits recorded actually appear in the aggregate genotype, but both can be expected to be closely related to the traits in the aggregate genotype.

6.1

Predicting Response to Selection and Related Parameters

If we were interested in a practical problem of what weight to give a series of observations on a particular population of animals for selection, and we were certain of our phenotypic and genetic parameters contained in P and G and were sure of our economic weights in v, then we might stop here. But curiosity alone would likely prompt us to ask what the predicted variance of the index would be, what would be the predicted response of the aggregate genotype to selection on the index, and what would be the predicted correlated response of each of the traits in the aggregate genotype? And, mere curiosity aside, such predictions are essential when comparing different possible indexes and in assessing whether predicted responses of individual traits are likely to be acceptable to the users of the index or to their customers. In section 4.4.1.2 the variance of the index, 2 I , was derived as

2 I = b'Pb

(6.4)

It is important to realize that this equation holds whatever values of b are used, i.e. not only for the optimal b derived using selection index theory, but for any arbitrary vector b. However, if b is the optimal set of index coefficients, then b=P-1Gv which when substituted into (6.4) gives,
-1 2 I = b'P P Gv = b' Gv

(6.5)

This second form is most often quoted as the variance of a selection index. But it should always be remembered that this is true only of the "optimal'' selection index. Since you may wish to explore the consequences of using sub-optimal indexes, it is probably safer to use the form given at (6.4), which holds for any selection index, optimal or sub-optimal.

6-3

Applying the general equation for response to selection, to selection on I to improve H, predicted genetic superiority for the aggregate genotype, H, to selection on the index, I, is given by SH = i rHI H (6.6) As given in 4.4.1.2, the variance of the breeding goal can be derived as: 2 H = v'Cv (6.7) where C is an n x n matrix of genetic covariances among the traits in the aggregate genotype. Similarly, it follows that HI = b'Gv (6.8) Hence, rHI =

HI = IH
b' Gv b' Pb

b' Gv b' Pb v' Cv

(6.9)

and

SH = i

(6.10)

Note again, all the above derivations, (6.6) to (6.10), apply to any selection index (that is any set of b values) not just the optimum index. However, if we are dealing with the optimum index, where b = P-1Gv, then as noted at equations (6.4) and (6.5) b'Pb = b'Gv and substituting into (6.11) equation (6.10) gives: SH = i b' Gv We could have also obtained this optimum selection response directly, as SH = i bHI I and recalling that for optimum selection index coefficients we did set bHI = 1, hence SH = i I and substituting in the variance of the optimum index defined at (6.5): SH = i b' Gv

(6.12)

The accuracy of an index is defined as its correlation with the aggregate genotype, i.e. rHI as estimated at (2.17). For an optimum index, noting the equivalence of equations (2.13) and (2.14), the expectation for the accuracy of the index can also be written as rHI =

b' Gv v' Cv

(6.13)

The expected change in the additive genetic value of the ith trait in the aggregate genotype due to selection on the index, Sgi, can be found as Sgi = i bgiI I = i
Where
gI I
i

(6.14) (6.15)

g I = cov( g i , b' x) = b' cov( g i , x) = b' G i

i

where Gi denotes the ith column of G:

6-4

G = [G1, G2, , Gi, , Gn] Hence, Sg i = i b' G i b' Pb (6.16)

for all indexes. For an optimum index, the solution is also b' G i Sg i = i b' Gv

(6.17)

It also follows directly from (6.17), that the vector of genetic responses of each of the traits in the aggregate genotype, can be found as b' G Sg = [ S g1 , K, S g i , K, S g n ] = i (6.18) b' Pb These derivations of the principal parameters defining a selection index for all indexes and for optimum indexes are summarized in Table 6.1 Table 6.1 Summary of selection index formulae for any index and for optimum indexes. Derivation Any Index Optimum Index Arbitrary P-1Gv b'Pb b'Gv v'Cv b'Gv b' Gv b' Pb v' Cv b' Gv i b' Pb b' G i i b' Pb b' G i b' Pb i v'Cv b'Gv b' Gv = I v' Cv H
b' Gv = i I

Parameters b

2 I 2 H HI
rHI SH Sgi Sg

i i

b' G i b' Gv b' G b' Gv

The correlation between index values of relatives, which is needed to, e.g. compute selection b' Rb intensities as in section 3.6, can be computed as: t = where R is a matrix with covariances b' Pb between information sources on the two relatives (De Boer and Van Arendonk, 1989).

6-5

6.2 Example of Economic Index Selection

A beef breed is to be used as a terminal sire and economic analysis has shown that three key traits are post-weaning gain (PWG) with an economic value of 370 $/kg/d, ultrasonic back fat depth (BF), with an economic value under the local payment system, of -20 $/mm, and feed intake, with an economic value of 50 $/kg/d. Only PWG and BF are recorded in the bulltesting program. Genetic and economic parameters are summarized in Figure 6.1.

Figure 6.1. Example Derivation of Economic Index Beef Cattle Terminal Sire Line
Post Weaning Gain (PWG) Back Fat Feed Intake

Econ value h2 p rp BF FI rg

370 $/kg/d 0.40 0.23 kg/d PWG 0.18 0.23

(BF) -20 $/mm 0.44 0.15 mm

(FI) -50 $/kg/d 0.20 0.5 kg/d Select on index of own phenotype for PWG + BF

BF 0.32 0.40

FI 0.70 0.48

PWG

Matrices P, G, and C have the following elements:

2 p PWG P = pPWG , pBF

PWG , p BF

2 p BF
PWG , g BF

(6.19)

pPWG , g PWG G= pBF , g PWG

2 g PWG C = g BF , g PWG g FI , g PWG

p
g

BF

, g BF

pBF , g FI
PWG , g FI

(6.20)

PWG , g BF

2 g

BF

FI

, g BF

g BF , g FI 2 g FI
PWG , g FI

(6.21)

The elements of P, G, and C can be found by recalling that

6-6

r12 =

12
2 2 1 2

hence and that

12 = r1212
2 2 g = h2 p

giving,

.0529 .00621 P = , .00621 .0225

.02116 G= .004632

.004632 .0099

.02277 .01068

and

.02116 .004632 .0099 C= .004632 .02277 .01068

.02277 .01068 .05

370 The vector of economic weights is: v = 20 50 123.6 Hence, the index weights would be b = P-1Gv = 9.52 The index accuracy would be
rHI =

b' Gv = 0.6214 v' Cv

The expected genetic superiority for the breeding goal to 1 standard deviation selection on the b' Gv SH = = 28.722 index is: b' Pb

Expected genetic superiority for PWG, BF, and FI to 1 standard deviation selection on the index .0926 b' G .0232 is Sg = = b' Gv .1015 Thus, 1 standard deviation of selection on the index would be expected to yield bulls with an average breeding value of +0.0926 kg/day for PWG, + 0.0232 mm for BF, and +.1015 kg/day more FI. Also, their expected breeding value for profit is +$28.722.

6-7

6.3 Economic Indexes Based on Estimated Breeding Values

(Dekkers) So far we have dealt with economic selection indexes that were based on phenotypic records. In these cases, the main interest is to predict the individual's breeding value for the aggregate genotype. In practice, however, the sources of information that are available to develop the economic index are EBV for individual traits, rather than phenotypic records:

I 1 + b2 g 2 + + bm g m = b g I = b1 g
This leads to a step-wise procedure for development of economic selection indexes, in which the first step consists of predicting BV for individual traits and the second step of combining the EBV into an economic index. An advantage of this step-wise approach is that it allows different breeders to put different emphasis on traits in the aggregate genotype, while utilizing the most accurate EBV for the component traits. Consider a vector of available phenotypic records x that can be subdivided into subvectors xi which correspond to phenotypic records on trait i for i=1,, m:
x = [x1, x2, x3, . . . . , xm]

and a breeding goal with n traits:

H = v1g1 + v2g2 + + vngn = vgH

Then, an economic selection index based on the full set of phenotypic records x can be derived as: IF = bFx with the vector of index weights derived using standard selection index theory from:
bF= P Gv
-1

In the two-step approach, the objective is to develop the following type of index:

I 1 + b2 g 2 + + bm g m = bS g IS = b1 g
Consider first the situation where the traits in the index and the breeding goal are the same. Thus, m=n and gH = gI = g . The first step is to derive EBV for each individual trait i based on all available data x:
i = bix g

The index weights bi can be derived as:

bi = P Gi

-1

where Gi is the ith column of matrix G, i.e. a vector representing the genetic covariances between the observations in x and the ith trait in the aggregate genotype gi.:

6-8

G = [G1; G2; ; Gn]

Because all EBV are based on the same data vector x, the selection index equations for the individual EBV can be combined to directly estimate the vector of EBV as:

= [g1, g2, , gn] = [b1; b2; ; bn]x g

Combining the selection index equations that are used to derive each set of index weights bi we get:

= {P [G1; G2; ; Gn]}x = (P G)x g

-1

-1

Using these EBV to develop the economic index we get:

= bS(P G)x = (P G bS)x IS = bS g

Note that, if bS = v then IS = (P-1Gv)x = bFx = IF. This equality proves that the optimal index that is based on EBV can be obtained by weighting the EBV with the economic weight for that trait and summing over the traits in the aggregate genotype. Note, however, that this only holds if the same traits are included in the breeding goal and the index and the EBV are based on all available data on all traits, i.e. multiple-trait EBV are used. If the traits in the breeding goal and index are not the same, i.e. m = / n and gH = / gI then the index weights can be derived by partitioning the breeding goal into a component that is related to the traits in the index and an uncorrelated residual:
H = vgH = v( b g H g I gI + e)

-1

-1

'

where b g H g I is a vector of regression coefficient for the regression of vector gH on gI . Vector

b g H g I can be derived using standard multiple regression methods as: b g H g I = C I-1 C IH where C I is the genetic variance/covariance matrix among the traits which appear in the index and C IH is the genetic covariance matrix between traits in the index and traits in the aggregate
genotype. The residual e is by definition uncorrelated to traits in the index and does, therefore, not need to be considered when deriving the index. Thus, considering only the first term, the new breeding goal is one for which the assumption gH=gI holds but with a new set of economic values:
H* = v*gI = v b g H g I gI

'

6-9

Thus

v* = b g H g I v

can be derived as: and the optimal weights on the EBV for index IS = bS g
-1 bS = C I C IH v

An alternative derivation was given by Schneeberger et al. (1992). Note that when gH=gI then

C I = C IH and bS simplifies to the original result: bS = v.

It is interesting to note that the index weights for indexes that contain EBV depend only on the economic values and genetic parameters. They do not depend on the accuracy of the individual EBV. Thus, the same index can be used for all animals. It is important to realize, however, that these results only hold if all observations on all traits are used to estimate the BV for all traits. This would be the case when multiple-trait genetic evaluation models are used. The previous equations do not hold for EBV from single-trait
i . The reason is that evaluation models, for which only data on trait i are used to estimate g single-trait evaluation methods do not consider covariances between trait records, whereas covariances are considered in the multiple-trait evaluation methods. However, if the single trait EBV have high accuracy, index weights can still be approximated by the economic values. The reason is that with high accuracy, correlated trait information has little impact on EBV and singletrait EBV approximate multi-trait EBV.

If single trait EBV do not have high accuracy, approximate weights on EBV must be derived using selection index procedures as:
-1 bS= PS GSv

i where PS = matrix with (co-)variances among single trait EBV g

i with true BV in the breeding goal GS = matrix with covariances of single trait EBV g

Elements of PS and GS must now be approximated by specifying a vector xi of sources phenotypic

i . Standard selection index theory can then be used to derive each EBV as: records for each EBV g

i = bixi g

with

-1 bi = Pii Gii

where Pii is the (co-)variance matrix for records in xi and Gii is the vector of covariances of gi with xi. Then, the covariance between two EBV can be computed as:

6-10

g ,g = cov(bixi, bjxj) = bi cov(xi, xj)bj = bi Pij bj

i j

where Pij is the (co-)variance matrix between records in xi and xj and the covariance between and EBV and a true BV can be computed as:

g , g = cov(bixi, gj) = bi cov(xi, gj) = bi G ij

i j

where G ij is a vector with covariances of xi with gj . Note that in contrast to indexes based on multi-trait EBV, weights on single-trait EBV depend on the accuracy of the EBV and on the sources of information that contribute to each animal's EBV and will, therefore, differ for each individual animal. Depending on parameters and the range of accuracies, it may however be possible to use a single index for all animals, with the index derived based on average amounts of information. Methods to evaluate the loss in accuracy when using a single index will be described in the next section. One issue that we have not yet discussed is how to derive the accuracy of economic selection indexes that are based on EBV. When the index is based on single-trait EBV, accuracy can be derived using standard selection index theory as:
rHI = b S' G S v v' Cv

For the index based on multi-trait EBV, although index weights do not depend on the amount of information that goes into each EBV, the accuracy of the index does. For such indexes, accuracy must be derived by specifying the sources of information in the multiple-trait vector of observations x. Then, accuracy can be obtained by deriving the accuracy of the full index:
rHI =
' bF Gv v' Cv

This is illustrated in the following example where selection is based on observations in growing pigs for growth rate (GR, g/d) on the individual (xGR) and the mean of 5 full sibs (for xGR ) and own performance for feed intake (xFI , g/d). The phenotypic standard deviation is 100 for GR and 200 for FI and the h2 for both traits is 0.25. The phenotypic correlation is 0.6 while the genetic correlation is 0.8. The aggregate genotype is H = 0.2 gGR - 0.05 gFI . The selection index weights for the prediction of the aggregate genotype using the full index with all three sources of information (xGR , xGR , xFI ) can be calculated as:

6-11

1000 0 1250 12000 bF = P-1Gv = 1250 3000 2000 12000 2000 40000

2500 4000 0.02874 0 . 2 1250 2000 0.05 = 0.04011 0.00313 4000 10000

The weights for the prediction of the breeding value for growth rate b1 using all data are:
10000 1250 12000 -1 b1 = P G1 = 1250 3000 2000 12000 2000 40000 1 2500 0.1702 1250 = 0.3239 0 . 0327 4000

Similarly the weights for feed intake using all data, b2, are:
10000 1250 12000 -1 b2 = P G2 = 1250 3000 2000 12000 2000 40000 1 4000 0.1061 2000 = 0.4935 0 . 1935 10000

As demonstrated previously, the multiple-trait EBV can be combined into an economic index to predict the aggregate genotype with index weights equal to the economic values: FI GR -0.05 g IS = 0.2 g Substituting the multiple-trait index weights b1 and b2 gives:

0.1702 IS = 0.2 0.3239 0.0327

' x

0.1061 x - 0.05 0.4935 GR 0 . 1935 x FI

GR

' x

0.02874 x = 0.04011 GR 0 . 00312 x FI

GR

' x

xGR GR x = b x F GR GR x FI x FI

proving that the index based on multiple-trait EBV is equivalent to the full index. Let us now look at a situation where breeding values for GR and FI are predicted with a univariate model, i.e. the breeding value for GR is based on observations for GR only and similarly the breeding value for FI is based entirely on observations for FI. The following index weights result:

10000 1250 b1 = P G11 = 1250 3000

-1 11

2500 0.2088 1250 = 0.3297

Since the single-trait EBV for FI is based only on own performance, the index weight is equal to heritability:

6-12

b2 = P22 G22 = [40000] [10000] = [0.25]

-1

-1

Combining these single-trait EBV into an economic index to predict H results in:
2 g g GR , g FI bS= P GSv = GR 2 , g g g GR GR FI

-1 S

g GR , g GR g GR , g FI v g , g g , g FI FI FI GR 1 b' P b b' P b b'G b'G 1 11 1 1 12 2 1 11 1 12 v = b' P b b' P b b' G b' G 2 21 1 2 22 2 2 21 2 22 1 934.184 791.25 934.125 1494.6 0.2 0.1292 = = 2500 791.25 2500 1000 0.05 0.0109

These weights on single-trait EBV differ from those on multiple-trait EBV, which are equal to the economic values. Multiplying the weights out gives:
0.2088 IS = 0.1292 0.3297 0

' x

0 x - 0.0109 0 GR 0 . 25 x FI
GR

' x

0.02698 x = 0.04261 GR 0 . 00272 x FI

GR

' x

xGR GR x *' GR = b F xGR xFI x FI

These weights are different from those for the optimal index and, thus, the index derived from single-trait EBV is not optimal. The reason for the suboptimality is that, when deriving the single-trait EBV, relative weights put on records are determined while only considering a single trait and not the whole aggregate genotype. In our present example, this resulted in a weight of 0.2088 on xGR and 0.3297 on xGR , a ratio of 0.2088/0.3297=0.633. This same ratio of weights is still present in the overall index (0.02698/0.04261=0.633). In the optimal index, however, the ratio of weights on xGR versus xGR is 0.02874/0.04011=0.7165. Although the single-trait EBV index is not optimal, the next question is how much we would lose in accuracy when using this instead of the optimal index. This can be studied by evaluating the accuracy of both indexes. Accuracy of the multiple-trait EBV index must be derived from the full index with the three sources of information: b F ' Gv rHI = = 0.5518 v' Cv The variance of the aggregate genotype vCv equals 45 in this example. The accuracy of the single-trait EBV index can be derived in two ways:

6-13

1) as the accuracy of a sub-optimal index with the three information sources: b* F ' Gv rHI = * b* F ' Pb F v' Cv 2) or as the accuracy of an optimal index based on the two single-trait EBV bS 'G S v rHI = v' Cv Both result in accuracy to be equal to 0.5511, which is only slightly smaller than the accuracy of the optimal index, which was 0.5518. Thus, in this case, use of single-trait EBV had limited impact on efficiency. As an example of the use of principles similar to those outlined above, Veerkamp et al. (1995) investigated the sensitivity to economic values and genetic parameters of a sire selection index of type and production traits for a breeding goal with production traits and longevity. They also investigated the efficiency of using index weights derived for an index with multiple-trait EBV in indexes that included single-trait rather than multiple-trait EBV and in indexes that included EBV from two multiple-trait evaluations, one for production traits and one for type traits. Results showed that losses in efficiency were less than 1% for EBV based on at least 50 daughter records, which confirms the robustness of selection indexes for progeny-tested dairy sires.

6.4 Subindex Approach to Development of Economic Selection Indexes (Dekkers)

When the breeding goal and selection index involve a large number of traits, a two-step approach can be used for development of an overall index. In this approach, which was described by Dekkers (1995), traits in the breeding goal are subdivided into subsets or subgoals, (e.g. for dairy cattle, a subgoal for production, which could include milk, fat, and protein yield, and subgoals for reproduction, and health or longevity). Corresponding to each subgoal, subsets of selection index traits are identified which are closely related to traits in the subgoal. The first step in development of the overall selection index then is derivation of a subindex corresponding to each subgoal. The second step involves combining the subindexes into an overall index. Selection index procedures can be used to derive each subindex, as well as to combine subindexes into an overall index, and are described in the following. Let the breeding goal comprise n traits, which can be subdivided into q sections or subgoals, in which each subgoal is composed of biologically related traits (e.g., production traits versus herd life traits versus udder health traits). Then, the overall breeding goal:
H = v1g1 + v2g2 vngn

Can be written in terms of subgoals as:

H = w1 HSUB(1) + w2 HSUB(2) + . . . + wq HSUB(q)

6-14

with

' HSUB(i) = v SUB ( i ) gSUB(i)

where gSUB(i) = vector with breeding values for traits in subgoal i, vSUB(i) = vector with relative economic values among traits in subgoal i, and wi = economic value of subgoal i relative to other subgoals. Multiplying wi and vSUB(i) gives the economic values in the original breeding goal, vi. Similarly, EBV for m traits in the index I can be separated into q subindexes, which correspond to the subgoals I = d1 ISUB(1) + d2 ISUB(2) + . . . + dq ISUB(q) with
SUB ( i ) where g bSUB(i) di

' ISUB(i) = b SUB ( i ) g SUB ( i )

= vector with EBV for traits associated with traits in subgoal i, = vector with relative index weights on EBV in subindex i, and = weight on subindex i relative to other subindexes.

A trait can appear in more than one subindex. Subindex weights bSUB(i) can be derived by applying methods described in section 6.1 and 6.3 to subgoal HSUB(i) , depending on whether EBV are based on single- or multiple-trait models. The vector of weights to be applied to each subindex (d=[di]) can be derived based on
' P b )-1 b ' Gv d = ( b SUB SUB SUB

where b SUB
P G v

= m x q matrix, of which column i represents subindex weights bSUB(i) augmented with zeros for selection index traits not included in subindex i, = m x m matrix of (co)variances among the m EBV included in the overall index, = m x n matrix of covariances between the m EBV in the index and the n BV in the breeding goal, and = n x 1 vector of economic values in the original breeding goal (v' = [wi v ' ]).
SUB ( i )

The previous equation holds for single-trait EBV. If accuracy of the individual subindexes is high, or genetic correlations between subindexes are small, weights di can be approximated by wi, as discussed in section 6.3. The accuracy of the resulting overall index (rHI) can be derived from
d 'b ' SUB P b SUB d v' Cv

rHI =

6-15

where C = n x n matrix of genetic (co)variances among traits in the overall breeding goal. The subindex approach allows development, interpretation, and explanation of selection indexes in a sequential approach. The use of subindexes has the following advantages, some of which were discussed by Wilmink (1996), who referred to subindexes as supertraits: 1) Subindexes summarize EBV for related traits into a smaller number of EBV for traits or trait complexes that are of more direct relevance to dairy farmers; 2) With limited knowledge on the economic and genetic parameters that are needed for development of an overall index, subindexes use parameters among biologically related traits, which tend to be known with more certainty; 3) Because of the tendency of traits within a subindex and subgoal to be more highly correlated, weights on traits within a subindex are more robust to errors in genetic and economic parameters; 4) Use of subindexes facilitates interpretation of weights in an overall index because weights on subindexes reflect more closely the emphasis on economic traits or subsets of economic traits in the breeding goal. This is especially the case because correlations between traits in different subgoals or subindexes are generally lower than correlation among traits within a subgoal and subindex; 5) Use of subindexes facilitates use of a mixture of EBV from single- and multiple-trait evaluations, where multiple-trait evaluations are often conducted on sets of traits that correspond to a subgoal or subindex; 6) Use of subindexes facilitates implementation of customized selection indexes because economic parameters are needed for only a limited number of traits or trait complexes; 7) With use of subindexes, new information or new traits can often be incorporated at the level of the subindexes without the need to modify to overall index. A disadvantage of the use of subindexes in development of an overall index is that subindexes may not result in the optimal index for the defined breeding goal. This is because weights among traits in a given subindex are derived independent of traits in other subindexes and subgoals and, thereby, ignore genetic correlations with traits in other subindexes and subgoals. The impact of the use of subindexes on accuracy of the overall index was studied by Kulak et al. (1996) who found that reductions in accuracy were small unless genetic correlations between traits in one subindex and traits in another subindex differed substantially between pairs of traits. This would, however, not be the case in most situations in practise. In fact, in most cases, the division of traits into subgoals and subindexes would be such that correlations between subgoals and between subindexes will be limited. In addition, lack of information on genetic and economic parameters between different subindexes and subgoals will limit superiority of the optimum index for a given set of parameters. Wilmink (1996) did not advocate combining subindexes into an overall index at the population level, apart from the use of subindexes in development of customised indexes. Although the availability of subindexes reduces the need for an overall index, we think that availability of an overall index is desirable to promote the proper use of subindexes in an overall genetic improvement strategy and to set an overall goal for genetic improvement in the population.

6-16

The use of subindexes formed the basis of development of the total economic value (TEV) index that was developed for dairy bulls in Canada (Dekkers, 1995). The TEV consists of subindexes for production (PROD), herd life (HL), and udder health (UH): TEV = 26 ( 10 IPROD + 4 IHL + 1.5 IUH ) with

P )/ gP + 2( g F )/ gF ]/11 P - g F - g IPROD = [9( g

IHL IUH

HL )/ gHL HL - g = (g
SCS )/ gSCS SCS - g = -( g

i = mean genetic evaluation for trait i of sires that constitute the genetic base for the TEV; where g = genetic standard deviation for trait i, P = protein yield, and F = fat yield. The subindex for production was developed for a subgoal with milk, fat, protein, and lactose yield, based on research by Gibson et al. (1992) and J. P. Gibson, M. Greimel and J. C. M. Dekkers (1995, unpublished). The subindex for herd life relates to the trait herd life in the breeding goal and is a published genetic evaluation for herd life that is based on survival of daughters in the first three lactations and an index of four conformation traits (Jairath et al. 1998). The subindex for udder health was based on a subgoal that included clinical and subclinical mastitis in first and later lactations (four traits). The udder health index initially included only published genetic evaluations for SCS. All EBV and weights in the TEV index and subindexes are standardized by the respective genetic standard deviations and the mean EBV.
i

6.5 Sensitivity of Selection Indexes to Estimates of Variances and Covariances

As described so far, the selection index provides a method to maximize selection response for a given aggregate genotype when a given set of observations are available. It is assumed that the variances and covariances that make up the elements of P, G and C are known without error. In practice, elements of P, G and C are estimated with error, and are often obtained from several different sources. Sales and Hill (1976) deal in detail with the ways in which errors in the estimates of variances and covariances affect the efficiency of selection. For the present purposes it is sufficient to note that, when the elements of P and G are not known without error, the selection index is often relatively insensitive to errors in the estimates of these elements. However, it is always wise to examine how sensitive the resulting index is to the assumed elements of P and G. (Note that the elements of C do not affect the selection index coefficients nor the prediction of response to selection.) Bouchez and Goffinet (1990) propose a method for eliminating traits (or other sources of information) from the index based on maximizing an approximate estimate of the mean square error of prediction of the index, allowing for the inaccuracy of estimates of variances and 6-17

covariances among traits. Their method currently deals with the situation where parameters are estimated from the same population as is being selected. Their method does not appear to increase selection response but involving fewer traits, is claimed to be more robust than a full index. An alternative approach would be to use Bayesian methods to simultaneously derive genetic parameters and selection indexes. Bayesian methods allow uncertainty about parameters to be included in the development of the index. Literature estimates of genetic parameters can be included through specification of proper priors. Bayesian analyses are, however, typically complex in nature and an alternative series of simpler tests and modifications is proposed here.

6.5.1 Testing Consistency of Variance/Covariance Matrixes

Whenever a selection index is constructed, the P and C matrixes should be tested for being positive definite. A non-positive definite variance/covariance matrix may indicate that one or more traits in the index are linear functions of combinations of other traits, or that certain elements of the matrix are not possible given that the other elements are correct. For example, consider the phenotypic covariances among three traits with standardized variances such that,
2 2 2 1 = 2 = 3 = 1

Then, if 12 = 0.9 and 13 = 0.9, common sense would tell us that 23 must also take a relatively high value. A value of 23 = -0.5 would clearly be impossible. It can be shown in this case that .9 .9 1 1 23 if P = .9 . 9 1 23
P becomes non-positive definite when 23 0.62. At this point P also becomes singular and an inverse to P (i.e. P-1) cannot be found.

All matrix programming languages carry sub-routines to calculate the eigenvalues of a matrix. There are as many eigenvalues as there are rows and columns of the matrix (i.e. an n x n matrix has n eigenvalues). If one or more of the eigenvalues is negative, the matrix is non-positive definite. For example, the above P matrix could be tested in SAS IML by a program,
proc IML; P = {1 .9 .9, .9 1 .62, .9 .62 1; x = eigval(P); print x;

This program would yield the output,

x 2.62 .38 -1.41 E-19

6-18

This shows that the third eigenvalue of P is negative and that P is non-positive definite. Matrices which are close to being non-positive definite should also be examined closely because such matrixes are very close to being singular and are only just within the allowable parameter space indicating possible inconsistencies in the parameter values assumed. As noted by Sales and Hill (1976) and Hayes and Hill (1980), matrixes are more likely to become inconsistent (non-positive definite) as the number of traits considered increases, the size of the data base used to obtain estimates decreases and when estimates are obtained from different sources. This in turn can lead to increasing instability and uncertainty over the resulting indexes and their responses. In general, this argues for keeping the number of traits in both the index and the aggregate genotype down to the minimum number consistent with achieving effective genetic progress. The P and C matrixes referred to above are those for the observations in the index and the traits in the aggregate genotype respectively. Often the elements of P are themselves derived from population phenotypic and genetic parameters. For example, the observations in the selection index might be full-sib mean performance for a series of traits. Then the variances and covariances among these observations would be derived from the phenotypic and genetic variances and covariances among single phenotypic records. We can denote PO and GO as the phenotypic and genetic variance covariance matrixes among single phenotypic records of traits appearing in the selection index, which we used to derive the elements of P. Then, if our records are to be biologically meaningful, PO and GO should also be within the allowable parameter space and should thus be tested for being positive-definite. Similarly, if our model is then
y=g+e EO = PO - GO

is the implied environmental variance/covariance matrix among the traits in the index given PO and GO. Again, to be consistent and biologically meaningful, EO should also be positive definite. Variance/covariance matrixes can be biologically and economically inconsistent despite being positive definite. For example, a matrix for the three traits, milk yield (M), fat yield (F) and fat concentration (f) could be positive definite but not conform to the expectation that, since F f = , the variance of f and its covariances with F and M are entirely dependent on the M variances and covariances among F and M. This is because the test of positive definiteness examines possible linear combinations of traits and does not recognize the possibility of there being a pre-specified linear or non-linear relationship, such as f being a ratio of F/M. In general, any trait that is a direct function of other traits in the index (or the aggregate genotype) will carry little or no additional information and as such is usually best omitted. The economic weight of a trait is defined as the marginal change in economic value given a genetic change in that trait while holding all other traits in the aggregate genotype constant.

6-19

Thus any trait in the aggregate genotype which is entirely a function of other traits in the aggregate genotype has an indeterminate economic value since it is impossible to make genetic change in that trait without making genetic change in other traits. As an example, again F , it is quite clear that there cannot be a change in f considering the traits M, F and f, where f = M without a change in either F or M or both. Thus, if F and M appear in the aggregate genotype, f should not appear. It is surprising how often these basic checks for consistency of variance/covariance, components are not made in examples appearing in the literature, sometimes causing quite misleading results and conclusions. It is thus best not only to check over your own parameters for consistency, but also to check parameters in published papers, unless the authors have quite clearly stated that the appropriate checks have been made. In many cases, by elimination of unnecessary and less useful traits in the index, matrixes will become positive definite. Where matrixes remain non-positive definite, Hayes and Hill (1981) present a mathematical method ("bending'') for altering parameters to obtain positive definite matrixes. This method assumes there is no additional information available about parameter estimates and as such is unlikely to be generally appropriate. An alternative is to adjust individual parameter estimates based on an assessment of reliability of the estimates. An example and further discussion is given in the following section. But there is certainly room for further investigation of this problem.

6.5.2 Tests of Sensitivity and Covariances

of

Indexes

to

Assumed

Variances

If the various variance/covariance matrixes pass the test of being positive definite and being consistent in other ways, there is still a concern that the index derived may be sensitive to the assumed values of the variance/covariance components. A test of sensitivity would be to ask what proportion of the maximum response would we expect in the aggregate genotype if we used one set of variances to derive our index coefficients when another set of covariances was the correct set. If we use the subscript u to describe the result of a particular set of parameters used and t to describe the set defined as true, then
bu = Pu-1 Gu v

and

bt = Pt-1 Gt v.

If Pt and Gt are the true variance/covariance matrixes, then, by definition, selection using index coefficients, bt, will give the maximum expected response in the aggregate genotype H. Thus use of bu will give less response than bt. We can then define the efficiency of using bu instead of

6-20

bt as the ratio of genetic superiority of selected individuals in H when selecting using bu SHu compared to when selecting using bt, i.e. Eut = SHt
S H u can be found from the equation in Table 6.1 that defines response to selection from an

arbitrary index and S H t from the equation which defines response to selection as from optimum index, so that
S Hu = i

b' u G t v b' u P t b u

and

S H t = i b' t G t v

and thus

Eut =

b' u G t v b' u P t b u

1 b' t G t v

The process would be to define a parameter set to be our best possible estimate based on the evidence available and to use this as Gt and Pt. Alternative sets would then be defined and the question asked of each alternative set, "if the alternative set were indeed true, how well would the index based on our best estimates perform?'' There will likely be some difficulty in deciding what are realistic alternative sets of parameters. Clearly if parameters are well estimated there would be little relevance in examining widely different parameter sets. Whereas, if we are very uncertain of our estimated parameter set, we may well wish to examine widely different alternative sets. Also, the problem becomes worse the larger the number of elements in P and G since the number of possible combinations of alternative parameter sets rapidly expands with increasing dimensions of P and G. If there is insufficient evidence to say that the parameters used are the "best available'' or "most likely'', it may be worth considering an alternative set of parameters if one can be found that gives an index which is less sensitive to the parameters being incorrect than use of the original set of parameters. This could be judged by creating a matrix of relative efficiencies where each parameter set was tested against all other parameter sets, i.e.
true parameters E 1n E 12 E 11 E E 22 E 2n 21 E= E n2 E nn E n1

used parameters

where the rows correspond to parameter sets used and the columns to the parameter sets assumed S Hu i to be correct, so that: Eij = S Ht
j

6-21

where i is the used set of parameters and j is the true set of parameters. The diagonal elements of E are 1.0, all other elements are <1.0. The parameter set could then be chosen for which the elements of the corresponding row of E were as close to 1.0 as possible, i.e. the parameter set which showed the least sensitivity to being incorrect in relation to the alternative parameter sets examined. These processes of investigation are rather subjective. Indeed there is scope for research into the general area of how best to obtain estimates of population parameters, how to perform sensitivity analyses and how to modify parameters for particular applications in animal breeding. Bayesian methods provide opportunities to address these issues. There is also the question of what action to take if our results are indeed sensitive to the parameter values and what constitutes sensitivity? For example, if our best estimated parameter sets give response efficiencies Eu in the range of 0.85-1.0 for alternative parameter sets, would this be considered as sensitive or insensitive? There is no hard and fast rule, though by analogy with advances in methods of genetic evaluation (which improve accuracy of evaluation and hence response), potential losses in response of one or two percent may be of little concern whereas losses of 5% or more at the very least warrant careful investigation.

6.5.2.1

A Numerical Example of Testing Variance/Covariance Estimates

Index

Sensitivity

to

Imagine a species such as swine or beef cattle, selected for growth performance, perhaps as a terminal sire. Assume that the traits in the aggregate genotype have been determined to be slaughter weight (SW) and feed intake during growth to slaughter (FI), and both these traits are recorded and a phenotypic selection index is used. When both traits are scaled so that p = 1, the economic weights of SW and FI were found to be 2 and -1 monetary units per phenotypic s.d. (Note that FI has a negative economic weight reflecting that feed costs money; increased FI decreases profit.) The heritability of SW and FI have been estimated from several data sets and found to be 0.4 for both traits. The phenotypic and genetic correlations have been estimated only once and found to be 0.8 and 0.8. Thus, since selection is on phenotypic records of the 1 .8 .4 .32 2 individual, P= , G= and v= . .4 .8 1 .32 1 gij Note that the off-diagonal element of G, gSW.FI , is obtained by recalling that r g ij = , 2 g2 gj i hence

SW.FI

= 0.8 0.4 0.4.

.8 From these parameters and selection intensity i we obtain, b = .4

and

SH = 0.5367 i .

6-22

In the present case we might decide that the heritabilities are well estimated, coming as they do from several published data sets, but we are uncertain about the estimated correlations. In general, phenotypic correlations are much more accurately estimated than genetic correlations. So we conclude that we are most uncertain about our genetic correlation which we believe may in reality lie anywhere between 0.65 and 0.95. We then investigate the following four situations for the correlations, all other parameters remaining constant: 1. rgSW,FI = 0.80 and rpSW,FI = 0.8 2. rgSW,FI = 0.65 and rpSW,FI = 0.8 3. rgSW,FI = 0.95 and rpSW,FI = 0.8 4. rgSW,FI = 0.80 and rpSW,FI = 0.7. To calculate E21 we will need to determine the index weight corresponding to situation 2. The G .4 .26 matrix in that case is: G= .4 .26 yielding new estimates of
1.233 b= 0.867

The response to selection when using the index weights for situation 2 while the parameters for b' 2 G 1 v S H u ,t = i = 0.5122 i situation 1 are the true parameters is: 2 1 b' 2 P 1 b 2 The efficiency of using the alternative parameter set 1 while the original set 1 is true is estimated S H U ,t 2 1 as: E21 = = 0.5122/0.5367 = 0.9545 S Ht
1

Similarly, the efficiencies for all the other combinations of parameter sets can be calculated with 0.9545 0.8480 0.9738 1 0.9545 1 0.6513 0.8746 results in: E= 0.8480 0.6513 0 0.9258 0.9751 0.8646 0.9444 1 Two things are worth pointing out in these results. At the first place, the uncertainty over the true value of the genetic correlation between SW and FI seems to be important in terms of response to selection since an index using a genetic correlation of 0.8 is only 84.8% effective if rg is as high as 0.95 although it is 95.5% efficient if rg is as low as 0.65. We conclude that, in terms of finding an optimum index for this selection program, we need to have a more accurate estimate of the genetic correlation between SW and FI than we have at present. In particular, it 6-23

will be important to know whether this genetic correlation is higher than our current estimate of 0.8. A reduction of the phenotypic correlation from 0.8 to 0.7 has a smaller effect on the efficiency than a similar change in the genetic correlation. Secondly, the efficiency is symmetric about the diagonal for the first three-parameter sets but not for the last set. In the first three parameter sets G changed but P remains constant. For those cases it has been shown that the efficiency criteria is equal to the correlation between indices. Given that similarity, this symmetry is no surprise. Parameter set 4 resulted in a different P compared to the other three sets and as a consequence, for example, the efficiency E41 is not the same as E14. The predicted optimum economic response (SH) for the first three parameter sets corresponding to rg = 0.65, 0.8 and 0.95 were 0.749, 0.537 and 0.422 economic units per standard deviation of selection intensity. The predicted absolute economic response to selection increases by 77.5% as rg decreases from 0.95 to 0.65. From experience, it seems to be a general phenomenon that altering variances and covariances has a much larger effect on the prediction of absolute economic response to selection than on the efficiency of one index versus another. Since the prediction of absolute economic response is used in assessment of the economic cost-benefits of selection programs, having good estimates of variances and covariances will often be of even more importance for cost-benefit assessments than for the optimization of a particular program in terms of the index to be used.

6.5.3

Uncertainty Over Economic Weights

As will be discussed later, economic weights are rarely known without error. Indeed, there is often considerable uncertainty of what the true economic weight might be, arising from uncertainties of biological and management models used and uncertainty about the values of different traits in future production systems and markets. In many cases we will therefore have a "best estimate'' set of economic weights and several, perhaps many, alternatives covering alternative present and future scenarios. As with investigations of uncertain variances and covariances, we can carry out analogous investigations for uncertain economic weights. In this case the subscripts u and t refer to the economic weights used and those deemed to be "true''. Then the efficiency of a set of index weights derived for the used set of economic weights when compared to that for the ``true'' set of economic weights is given by, S Hu b' u G v t 1 Eut = = S Ht b' t G v t b' u P b u where and
bu = P-1 G vu bt = P-1 G vt.

6-24

As with uncertainties over variances and covariances, Eut could be derived for a single set of "most likely'' or "best estimate'' economic weights compared to a variety of alternative economic weights. Or, a matrix of efficiencies, E, could be determined comparing every set of economic weights with every other set. Economic weights do not affect P and, as shown in the previous section, E is symmetric about the diagonal so that Eij = Eji.

6.6 The Value of Including Traits in the Selection Index

Recording of animals takes time and effort and consequently costs n money. Some traits may be considerably more difficult and costly to record than others. Thus a key question in the design of breeding programs is which traits and types of relatives should be recorded? It is often relatively straightforward to identify a number of traits and types of relatives potentially useful to record. A selection index could then be constructed and the question asked, how much does each observation contribute to the response of the aggregate genotype? Then the economic benefits of including that observation in terms of enhanced response can be evaluated against the cost of recording and a decision taken on whether or not to collect that information. Consider the usual selection index problem with observations in the index and m traits in the aggregate genotype, but define the index as including all traits that might possibly be included. Then the efficiency of a reduced index without observation i can be defined as the ratio of economic (aggregate SH* i genotype) response for the reduced index to that with the full index, or: Ei = SH
* where S H is the response of the aggregate genotype for the index missing observation i, S H * , i
i

and SH is the response with the full index I,

SH* = i
i

b*i G * i v = i I *
i

where b i is

* -1 b* G* i = Pi i v

and P *i and G *i are the reduced forms of P and G corresponding with the reduced index. P *i is found by deleting the ith row and column from P, and G *i by deleting the ith row from G. Similarly, the efficiency of indexes with more than one observation deleted can easily be found. SH* ij For an index with observations i and j missing: Eij = SH Where And

SH * =
ij

* * b ij G ij v

* * -1 * b ij = P ij G ij v

* * and P ij is P with the ith and jth rows and the ith and jth columns removed and G ij is G with the ith and jth rows removed.

6-25

The efficiency, given a certain aggregate genotype and set of population parameters, is directly related to changes in rHI. The contribution of each individual observation in the index to the accuracy can be calculated. Cunningham (1969) has described a method to derive the decrease in accuracy which is given below. Assume an index with n sources of information, I = b'x and an aggregate genotype with m traits, H = v'g. Solving the set of equations, Pb = Gv, gives the vector b. The variance of the index is b'Pb. When, for instance, the first observation is ignored, we obtain a new index, I 1* , with a vector with n-1 weighting factors, . The weighting factors are a solution of solving: P1* = * G1 v where, as before, P1* is the n-1 by n-1 variance-covariance matrix of the observations in I 1* , * and G 1 is the n-1 by m matrix obtained from G by ignoring the first row. Based on the original equation, this can be visualized as:
Pb = Gv

P11 Q b1 G1 = * v = * b Q P 0 G 1 1

R 1 R 0

where P11, b1 and R1 are scalars representing the first elements in P, Gv, and b, respectively. Q is a column vector representing the remaining elements of the first column in P. In most cases, b0. The variance of the reduced index I 1* is:

I2 = ' P1*
* 1

and its correlation with the aggregate genotype H is: rHI * =

1

' P1*
2 H

The efficiency of the index ignoring observation i is equal to E I * ,I =

i

SH*
i

SH

rH , I *
i

r H ,I

* i

, which

depends entirely on the difference in the variances of the indices, ' P1* and b'Pb: To determine the magnitude of this difference, let W be the inverse matrix of P: W11 ..............W1n ......................... -1 W=P = ......................... Wn1 ..............Wnn 2 b1 2 2 It can be shown (Cunningham, 1969) that: I - I1* = W11
In general, ignoring the i observation in the index: -
th
2 I 2 I i*

b i2 = Wii

6-26

The advantage of this method is that in computing the decrease in variance, no new index has to be derived; information for the computation is available from the computations for the original index. The variance of I2* can be derived from the equation above as:
1

b i2 Wii The efficiency of the index ignoring observation i can also be derived from the above results as:

I2 = I2
* i

EI * ,I =
i

SH*
i

SH

rH , I *
i

r H ,I

I2 I2

b i2 Wii

The method can be extended to compute the reduction of variance of the index, simultaneously ignoring more traits. The variance of an index, from which the ith until the jth observation is 1 ' ignored, is I2* = I2 b 'i..j Wi..j b i..j
i .. j

where bi..j is the vector including ith until the jth weighting factor of the original index and Wi..j is the diagonal submatrix of P-1 corresponding with the removed observations from the index.

6.6.1

Example of A Reduced Index

Consider the example in section 6.5.2.1 for selection on slaughter weight (SW) and feed intake (FI). How much do inclusion of SW and FI in the index contribute to selection response? The 1 .8 .4 .32 2 , with v = full index was given by P= G= .4 .8 1 .32 1 Giving .8 b = , I2 = .288 , .4 and

SH = 0.537 i

An index without SW would be found from

P* 1 = [1], G* .4], 1 = [.32

and

2 v= 4

giving and hence

b* 1 = 0.24

and = 0.447

S H * = 0.24 i
1

E1 =

SH*
1

SH

Similarly, removing FI from the original index, gives

P* 2 = [1], G* .32] 2 = [.4
S H * = 0.48 i
2

and

2 v= 1

giving

b* 2 = 0.48,

6-27

and

E2 = 0.894.

The results could also have been obtained from b and P-1 which is equal to: 2.778 P-1 = W = 2.222 2.222 2.778 .288 .8 2 2.778 = 0.447 .288 (.4) 2 2.778 = 0.894 .288

Thus:

E1 =

2 I

2 b1 W11

I2

and:

E2 =

I2

b2 2 W22

I2

.288

Thus, not recording SW would reduce economic response by (1 - 0.447) 100 = 55.3% while not recording FI would reduce economic response by only 10.6% compared to recording both traits. Since recording FI is likely to be more expensive than SW, it would seem well worthwhile to ask in detail how much it will cost to record FI in the breeding program and what the expected extra 10.6% economic response is worth. The final answer will depend on such parameters as the number of animals in our breeding program, the number of animals benefiting from the extra genetic response, the cost of recording individual animals and the time between incurring costs (recording) and obtaining returns (selling or producing from genetically improved animals).

6.7 The Value of Traits in the Aggregate Genotype

With equation (6.11): SH = i b' Gv , the response in the aggregate genotype as a result of selection using the index can be calculated. The response to selection in the aggregate genotype is an important criterion in comparing different indexes. Besides this response it is also interesting to look at the response in the individual traits in the aggregate genotype. This will give information on which traits have contributed to the response in the aggregate genotype and one might want to look at the direction and size of change in each of the traits. In the example of beef cattle it is interesting to see whether the response to selection is due to a change in slaughter weight, feed intake or a combination of these two traits. In this section we will illustrate the relationship between the change in the aggregate genotype and the individual traits. Furthermore, we will look at comparing the response to selection for various situations and try to give some guidelines on the criteria to use in such a comparison. The expected change in the individual traits (gj) of the aggregate genotype as a result of selection on the selection index (I) can be computed using the covariance between gj and I as shown in

6-28

Recall that the expected change in gj (expressed in units of measurement) as a b' G j consequence of selection on I is: Sgj = i I th where Gj denotes the j column of G, and that the vector of responses in each trait is equal to: b' G Sg = i I The response to selection in the aggregate genotype SH in monetary units (assuming that elements of v are expressed in monetary units) can be calculated as: b' Gv SH = i I b' G j b' G which is equal to: SH = i v = i Sgv = i vj I I (6.18). The above equation demonstrates that the response to selection in the aggregate genotype is the sum of the response in the individual traits summed by their respective economic weights. The contribution of response in gj to the overall response (Cj) can be calculated as: Sg v j Cj = j SH With this criterion we can evaluate the relative contribution of each trait in the aggregate genotype. To illustrate this we return to the example of 6.5.2.1, use parameter set 1 and set the selection intensity equal to 1. We have seen that the response in the aggregate genotype SH equals 0.537. Using (6.18) the response in slaughter weight is 0.358 and the response in feed intake is 0.179, in both cases expressed in phenotypic s.d. The economic weight is 2 and -1 for SW and FI, respectively. Using these weights, the response in monetary units in SW and FI is equal to 0.716 and -0.179, respectively. The contribution of SW to the overall response (CSW) is equal to 0.716/0.537 = 1.33. In words, the response in SW accounts for 133% of the monetary response due to selection on the aggregate genotype. At first sight it looks strange that the contribution exceeds 100%. When we have a closer look at the example we see that we are trying to improve to traits which are positively correlated of which one has a positive and the other has a negative economic weight. Given the parameters we have used the index results in an increase in slaughter weight which is associated with a change in feed intake. The latter change results in a reduction in the financial gain. This is reflected by the negative contribution of 33% of FI to the overall response. From the example it is clear that Cj can take values which are smaller than zero or bigger than one. The advantage of looking at values for Cj instead of Sgj is particularly the unit used for the traits, i.e. money vs units of measurement. Different indexes can be compared using the accuracy of selection as long as the parameters for the aggregate genotype (e.g. genetic variance) and the economic weights have not changed. Comparison of indexes in other cases is less straightforward. Let us, for example, return to the example and consider a second index for a situation where feed intake was excluded from the aggregate genotype and two different values of rg, all other factors including the index remaining unaffected. Results for that and the original aggregate genotype are summarized in Table 6.2. 6-29

Let us first look at the situation where rg is 0.8. The accuracy of rHI was identical for both sets of economic weights but the variance of the aggregate genotype differed largely. We are looking at a situation with two highly correlated traits with opposite signs for their economic weights. Set the economic weight equal to zero for one of the traits which will, in that case, result in a higher variance of the aggregate genotype. The response to selection in the aggregate genotype and slaughter weight increased when v2 was set equal to zero. The correlated response in FI for v2 = 0 amounted to 0.32 phenotypic s.d which was much larger than for v2 = -1. Changing the economic weight of FI did not affect rHI. This is not a result that one would find in general but one which is specific for the parameter set used in this example. There are two factors that contributed to this. In the first place, both traits had the same heritability and standard deviation. Schaeffer (1984) showed that the benefit of including a correlated trait in the prediction of the EBV for a trait depends on the absolute difference between the phenotypic and genetic correlation. The correlated trait does not contribute when the phenotypic correlation is equal to the genetic correlation which is the case for rg = 0.8. This means that v2 = 0 refers to an index with observations on SW only for the correlation of 0.8. In that case one also expects to find rHI = 0.4 = 0.632. Different results are found when rg is lowered to 0.65. For v2 = -1, a positive response of 0.358 units in SW is associated with a reduction in FI of 0.035 units, i.e. a change in the desired direction for both traits in the aggregate genotype. As to be expected, the correlated response in FI is positive when v2 = 0 but it is smaller in size than when the genetic correlation was 0.8. In comparing the results in Table 6.2 it is obvious that there is not a single criterion which can be used. The criteria to look at very much depend on the question one wants to answer. It is important to realize that the variation in the aggregate genotype differs between all four cases studied. This is different from a situation where one varies the observations to be included in the 2 and can use the rHI. In comparing the index. In that case one generally works with a constant H four situations in Table 6.2, the efficiency criteria introduced in section 6.5.2 might be very useful. The efficiencies for the different combinations of parameters sets, using the numbering of sets as given in Table 6.2, can be summarized as: 1 0.954 E= 0.894 0.976 0.954 1 0.720 0.867 0.894 0.720 1 0.970 0.976 0.867 0.970 1

This illustrates that when rg is 0.8, the loss in selection response amounts to (1 - 0.954)*100% = 4.6% when FI was ignored in the aggregate genotype while the true economic value of FI is -1.

6-30

Table 6.2 Results of selection using a selection index including slaughter weight (SW) and feed intake (FI) for two different values of the genetic correlation between SW and FI rg and economic weight of FI v2 using a selection intensity of 1a). rg = 0.8 rg = 0.65 v2 = -1 v2 = 0 v2 = -1 v2 = 0 0.632 0.632 0.765 0.652 rHI 0.537 0.800 0.750 0.825 I 0.849 1.265 0.980 1.265 H 0.537 0,800 0.750 0.835 SH 0.358 0.4 0.358 0.412 SSW 0.179 0.32 -0.035 0.175 SFI 1.33 1 0.953 1 CSW -0.33 0 0.047 0 CFI Set 1 2 3 4 a The index in all cases consists of observations on FI and SW on the individual animal.

6.8

Non-Linear Indexes

The discussion of selection indexes so far has assumed that the aggregate genotype is a linear function of additive genetic values. In practice the assessment of net economic value as the aggregate genotype will often not be linear. In general, linear indexes can be found which closely approximate most non-linear descriptions of economic value. These will be discussed in relation to derivation of economic weights in Chapter 7.

6.9 Constrained and Desired Gains Indexes

(Dekkers)

In the selection indexes discussed so far, the response to selection of traits in the aggregate genotype is determined entirely by the economic weight of that trait, the information available in the index and the phenotypic and genetic variances and covariances among traits. It is, however, possible to construct indexes in which the rate of genetic change of one or more traits is in some way pre-determined. For example, it is possible to maximize an index for genetic change in one set of traits subject to one or more genetic traits being constrained to zero genetic change. Alternatively, the change in one trait might be desired to be twice the change in another, with all other traits being allowed to change as the index dictates. Methods for achieving these types of constrained or desired gain indexes are reviewed by Brascamp (1984). Linear programming (see later Chapter) can also be used to achieve constrained and desired gain indexes (Keller and Gibson, 1990; Toro, 1992).

6-31

Following Brascamp (1984), separate the vector of traits in the aggregate genotype H = vg into g 0 two sets of traits: g= g1 where g0 includes traits for which progress will be maximized according to v0g0 , i.e. the economic part of the breeding goal and g1 includes traits for which changes are constrained to relative changes = . G 0 G= Correspondingly, partition matrix G as: G 1 Let the constrained index be:

I*= b*X

with variance 2* = b*'Pb*

I

For simplicity, and without loss of generality, set 2* = 1

I

Then, from Table 6.1, response in g1 to selection on this index is proportional to b*G1 Thus, the problem of finding the constrained index can be formulated as the following constrained optimization problem:
Max b * ' G 0 v 0 b*

subject to b*G1 = and b*'Pb* = 1

This problem can be solved using the Lagrange multiplier method, as described in Brascamp (1984) and Weller (1994). Although there is considerable literature on the construction and properties of such indexes, at no point does there appear in the literature a sound reasoning for applying these indexes in economic animal breeding. Gibson and Kennedy (1991) attempted to provide a rationale for why such indexes should not be used in animal breeding on the grounds that an economic index can always be found to at least equal and usually to out-perform a constrained index. They showed that constrained or desired gains indexes have implied, or pseudo economic values for all traits, including for those that are constrained. The vector of pseudo economic values, v*, is the set of economic values that would have resulted in the index weights b*, using the selection index equations. The pseudo economic values can be derived as follows: Using the selection index equations, set
Pb*= Gv*

and solve for v* as follows:

GPb*= GGv* v* = (GG)-1GPb*

6-32

6.9.1 Example of Desired Gains Index

Figures 6.2 and 6.3 illustrate the potential impact of putting a constraint on genetic progress for a given trait. The example is of selection for an aggregate genotype in pigs consisting of growth, feed intake, and backfat thickness. Economic values are 0.178 $/g/d for growth, -0.05 $/g/d for feed intake, and 0.0415 $/mm for backfat thickness:

H = 0.178 ggrowth 0.05 gfeed intake 0.0415 gbackfat

The index includes phenotype on the individual and its sire for all three traits. Based on these economic values, unit intensity of selection on the optimal index results in an increase backfat thickness by 0.7316 mm (Figure 6.2). In many breeding programs, such an increase is undesirable and, thus it seems reasonable to develop a constrained index in which the genetic gain for backfat thickness is constrained to zero. The results of such an index are given in Figure 6.3. For this index, the relative pseudo economic values are: 0.02659 v* = (GG) GPb* = 0.007469 compared to v = 0.585
-1

0.178 0.05 for the optimal index. 0.0415

Note that economic values in v* are relative economic values. When expressed relative to an economic value of 0.178 for growth rate, which is its true economic value, implied economic 0.178 values under the constrained index become: v* = 0.05 3.916 Thus, the implied economic value for fat thickness relative to growth is almost 100 times as large as its true economic value. The efficiency of the constrained index in terms of the true aggregate genotype can be evaluated using the procedures described in sections 6.5.2 and 6.5.3. Alternatively, with genetic superiorities in the individual traits already computed for the desired gains index, response in H can be computed as described in section 6.7 as: 0.178 * * SH = Sg v = [13.65 , 67.37 , 0] 0.05 = $5.80 0.0415 In contrast, response for the optimal index is Thus, efficiency is:
S 5.80 E= H = = 0.785 7.3861 SH
*

SH = 7.3861
, a 21.5% loss.

(see Figure 6.2)

6-33

Figure 6.2 Selind program output for optimal economic index for pig selection example
Listing of used matrices -----------------------Listing of P-matrix 7744 968 1.144e+04 968 7744 220 1.144e+04 220 4e+04 220 1.144e+04 5000 47.52 27.74 56 27.74 47.52 29.66 Listing of G-matrix 1936 440 968 220 440 1e+04 220 5000 55.47 59.33 27.74 29.66 Listing of C-matrix 1936 440 440 1e+04 55.47 59.33

220 1.144e+04 5000 4e+04 29.66 56

47.52 27.74 56 29.66 4 1.1

27.74 47.52 29.66 56 1.1 4

55.47 27.74 59.33 29.66 2.2 1.1

55.47 59.33 2.2

Listing of P-inverse-matrix 0.0002572 -6.187e-05 -7.397e-05 -6.187e-05 0.0002572 2.736e-05 -7.397e-05 2.736e-05 4.751e-05 2.736e-05 -7.397e-05 -1.344e-05 -0.001678 -0.0004214 0.0001265 -0.0004214 -0.001678 -1.121e-05

2.736e-05 -7.397e-05 -1.344e-05 4.751e-05 -1.121e-05 0.0001265

-0.001678 -0.0004214 0.0001265 -1.121e-05 0.2886 -0.06352

-0.0004214 -0.001678 -1.121e-05 0.0001265 -0.06352 0.2886

Index based on economic weights =============================== ---------------------------------------------------| trait | ec weight | genetic sup | ---------------------------------------------------| growth | 0.178 | 30.88 | | feed intake | -0.05 | -38.39 | | fat thick | -0.0415 | 0.7316 | ---------------------------------------------------------------------------------------------------------------------------| animal , trait | B-value | rvi (%) | ------------------------------------------------------------------------| own performance , growth | 0.08516 | 30.5 | | sire , growth | 0.01686 | 1.018 | | own performance , feed intake | -0.03579 | 28.88 | | sire , feed intake | -0.007106 | 0.9787 | | own performance , fat thick | 1.095 | 3.882 | | sire , fat thick | 0.1251 | 0.04967 | ------------------------------------------------------------------------Standard deviation of index : 7.3861 of aggregate genotype : 8.8283 Correlation between index and aggregate genotype

: 0.83663

6-34

Figure 6.3 Selind program output for index for pig improvement example with gain for backfat thickness restricted to zero
RESULTS SECOND RUN WITH D AS DESIRED GAINS INPUT Listing of Gtranspose*P-inverse*G 1252 -136.6 38.44 -136.6 5162 27.62 38.44 27.62 1.395 Listing of S 0.05653 0.011 -1.776 0.011 0.002356 -0.3498 -1.776 -0.3498 56.6

Pseudo relative economic values based on: v* = (GG)-1GPb*

---------------------------------------------------| trait | ec weight | genetic sup | TRUE EC. VALUES ---------------------------------------------------| growth | 0.02659 | 13.65 | 0.178 | feed intake | -0.007469 | -67.37 | -0.05 | fat thick | -0.585 | 2.313e-15 | -0.0415 ---------------------------------------------------ECONOMIC RESPONSE = (0.178)*(13.65)+(-0.05)*((-67.37)+(-0.0415)*(0) = 5.80 Compared to 7.3861 for the optimal economic index (21.5% reduction!!) ------------------------------------------------------------------------| animal , trait | b-value | rvi (%) | ------------------------------------------------------------------------| own performance , growth | 0.009932 | 30.1 | | sire , growth | 0.002312 | 1.395 | | own performance , feed intake | -0.004965 | 44.49 | | sire , feed intake | -0.001173 | 1.95 | | own performance , fat thick | -0.1071 | 2.685 | | sire , fat thick | -0.04554 | 0.4801 | ------------------------------------------------------------------------Standard deviation of index : 0.86603 of aggregate genotype : 1.139 Correlation between index and aggregate genotype

: 0.76033

For H with implied economic values

6-35

6.10

Some Practical Considerations

(Taken from Dekkers and Gibson, 1997)

The objective of development of total merit selection indexes is to provide a guide to selection that maximizes selection for the overall breeding objective and to promote selection strategies that minimize misuse of information. Although development of total merit selection criteria can be based entirely on the scientific principles that were summarised above and that will be described further in Chapter 7 with regard to estimation of economic values, implementation of selection indexes in the industry requires careful consideration of acceptance of the index by its target group (i.e., the breeders). An index that is based on the application of sound scientific principles but is not accepted has much less impact on selection for an overall breeding goal than an index that may not be optimal technically but that receives acceptance by the targeted users. The main issues involved in the development of selection strategies from a technical versus a practical perspective are summarized and contrasted in Figure 6.2. To overcome these differences in perspective, development of selection indexes for use in the field must take place in close collaboration with the industry. Although this interaction may compromise scientific principles and objectivity, there are several aspects that can be pursued from a technical basis that can aid in development of an index that is technically sound while simultaneously maximizing acceptance of the index by its intended users. Several of these approaches, as well as their theoretical background, are discussed here (taken from Dekkers and Gibson, 1997), while others have been discussed earlier in this Chapter and in Chapter 7.
Figure 6.2 Factors affecting development of selection strategies technically and from a producer perspective.

Technical
ali sc e R ng

Breeding Goal Economic Values

Production economics

Selection Strategy
Quantitative genetics

Quota Production Restrictions System

On-farm prices & costs
t em sys rket a ota Qu ota m Qu

Future Markets
rep orts

InterGenetic national Parameters Trade

Profit

Producer

Selection and Mating Strategy

6.10.1

Development of Selection Criteria

A total merit selection index combines EBV (or ETA) for individual traits into a single number that can be used for selection. Although this index facilitates simultaneous selection for multiple traits, most breeders require insight into the index and, in particular, into the index weights. When 6-36

Me dia

derived based on selection index principles, weights in a selection index are equivalent to partial regression coefficients in a multiple regression equation. Similar to the difficulty of interpreting partial regression coefficients in a multiple regression analysis when explanatory variables are confounded, selection index weights can be difficult to interpret when traits are genetically correlated. This complicates extension of the index to producers. There are several avenues that can be pursued in a) derivation of the index, b) expression of the index, or c) extension, to facilitate the understanding and, thereby, implementation of the index by its target audience. Index derivation. Four approaches, among other, that can be used to develop total merit indexes that facilitates their understanding, acceptance, and implementation of the index are 1) alignment of index weights with economic values, 2) evaluation of the accuracy of alternative index weights, 3) use of customized indexes, and 4) development of indexes based on subindexes. In principle, all four approaches focus on consideration of alternative indexes that are more acceptable or easier to interpret and on exploring the accuracy of these indexes relative to the optimum index. Flexibility in consideration of alternative indexes without compromising accuracy stems from the robustness of genetic change to economic values, which was first explored by Vandepitte and Hazel (1977) and Smith (1983).
Aligning index weights with economic values. Economic values of traits in the breeding goal are frequently easier to interpret than are selection index weights. Therefore, aligning index weights closely to economic weights in the breeding goal will facilitate understanding of the index. The ability to accomplish this depends on the degree to which traits in the index differ from traits in the breeding goal, whether single- or multiple-trait procedures are used for genetic evaluation of traits in the index, and, in the case of single-trait evaluation procedures, on the accuracy of EBV.

For index traits that are indicators of traits in the breeding goal, index weights can be related to their indirect economic importance based on the relationship between the indicator trait and the economic trait in the breeding goal. For example, for SCC as an indicator of susceptibility to mastitis, an indirect economic value can be derived as the genetic regression coefficient of SCC on mastitis, multiplied by the economic value of mastitis. This derivation was implicit to the economic value of SCC derived by Schutz (1994). When EBV for traits in the index are based on a joint procedure for multiple-trait evaluation and traits in the index are the same as the traits in the breeding goal, economic values can be used directly as selection index weights, as described previously. When traits in the index are different from traits in the breeding goal, procedures described in previously can be used, and resulting index weights amount to the indirect economic values described. When EBV are from single-trait genetic evaluation models, index weights can still be approximated based on economic weights if the accuracy of individual EBV is high. The efficiency of indexes derived on this basis relative to the optimum index must be considered in these cases.
Exploring the accuracy of alternative index weights. Optimum index weights for a given set of economic values can be difficult to accept by producers if those weights do not correspond to the perceived incentives for genetic change, as discussed

6-37

previously. For example, Gibson et al. (1992) found optimum index weights for Ontario to result in a negative weighting for milk yield, although producers were paid in part for milk volume, in addition to payments for fat and protein yield. The negative weighting on milk yield was a result of the nature of the dual quota system that was in operation in Ontario at the time. In this sytem, one quota was based on volume for fluid milk sales, with a substantial premium for milk volume, and the second quota was based on fat sales, with no premium for milk volume. Because of the partial payment for volume, an index with a negative weighting on milk was difficult to accept by producers. An index with a zero weighting on milk volume was explored as an alternative. This index was found to be over 98.5% as efficient as the optimal index and was subsequently implemented by the industry as part of the LPI. Another aspect that helped implementation of this index was the demonstration that selecting on this index was expected to result in substantial improvement in milk yield, despite a zero emphasis on milk in the index.
Customised Indexes. Customized indexes allow producers to develop a selection index based on economic circumstances that are specific to their herd Bowman et al. (1996). The use of customized indexes is justified technically if economic circumstances differ between herds or if traits are genetically different across herds Visscher and Amer (1996), which would make use of a single index derived based on economic circumstances of an average farm inappropriate. Visscher and Amer (1996) found that, for economic and genetic parameters typical for dairy cattle, customized indexes did not result in substantially greater improvements in profit at the population level than a single index that was based on average parameters. This result was caused in particular by the dominating economic importance of production traits and the high genetic correlations among milk, fat, and protein production. Visscher and Amer (1996) concluded that the main reason for use of customized indexes in dairy cattle is promoting the use of a selection index approach to multipletrait selection instead of selection on independent culling levels. In development of customized indexes care must be taken, however, that objective economic information is used in their derivation; Visscher and Amer (1996) assumed that customized indexes were based on accurate economic data for individual farms.

Although customized indexes can play an important role in implementation of selection index principles in the field, we do advocate availability and promotion of an overall index that applies to average circumstances in the population. Such an index can provide default values for weights in customized indexes and provide a unified and global focus for selection and marketing decisions for all sectors of the breeding industry.
Development of indexes based on subindexes. This was described previously. 6.10.1 Expression of the index

Interpretation and implementation of an overall selection index can be facilitated by its expression. Expression of this index includes the name given to the index which must convey the purpose and meaning of the index, expression of the index formula, and the scale on which the index is expressed. The latter two issues are discussed further.

6-38

Expression of the index formula. The magnitude of economic values and index weights is dependent on the scale of expression of EBV and may not reflect the relative emphasis on traits in relation to genetic selection decisions. Multiplying index weights by the genetic standard deviation for the trait provides standardized weights that reflect the emphasis put on each trait in relation to the genetic variability that is present in the population.

Index weights can also be standardized by the standard deviation of EBV. Although the type of standardization (i.e., based on the standard deviation of true versus estimated breeding values) does not affect the eventual index values, there are some significant differences with regard to interpretation of the resulting index weights. These are summarised in the following: 1) index weights standardized by the genetic standard deviation are independent of the accuracy of genetic evaluations and are, therefore, more closely related to the economic importance of traits in the breeding goal; 2) standardization on the basis of the standard deviation of EBV reduces the emphasis that is perceived to be put on traits with low heritability (i.e., low accuracy of EBV). Standardization on the basis of the genetic standard deviation maintains the distinction between the ability to identify genetic differences in the population and the relative emphasis that is put on a given difference in EBV between animals; 3) the relative magnitude of the standard deviation of EBV among traits may differ between, for example, sires and cows. This would result in different indexes if standardization is on the basis of the standard deviation of EBV.
Scale of expression. A total merit index ranks sires and cows based on genetic merit for the overall breeding goal. Apart from use of the index as a ranking tool, understanding, implementation and use of the index can be enhanced if the index is expressed in meaningful units. For total merit indexes that are intended to rank animals based on genetic merit for profitability, expressing the index in monetary units facilitates and promotes use of the index. The ultimate goal is to express the index in a way that enables its use for investment decisions (e.g., semen purchase). Expression of total merit indexes for sires as a net present value of a dose of semen was explored by McGilliard (1978), Bakker et al. (1980), and others. This requires consideration of the time and frequency of expression of genetic superiority in resulting daughters, discount rate, and conception and survival rates. Based on such indexes, differences between bulls in expected returns from a dose of semen can be compared directly to differences in semen price. 6.10.2 Extension of the index

Effective implementation of any index relies on promotion of the index through extension activities. Extension efforts should focus on the consequences of selection on an index rather than on the composition of the index. Index weights or economic values may not give a clear indication of what can be expected from use of the index. This is because responses to selection on an index are affected by genetic constraints on improvement of individual traits, which are quantified by the genetic parameters, as well as by the emphasis put on each trait in the index through the index weights.

6-39

Responses to selection can also be used to illustrate that, although for dairy cattle an increased protein to fat ratio may seem a reasonable selection objective, and selection indexes can be derived that maximise that objective, selection on such an index may have detrimental effects on responses for yield traits and result in reduced fat and protein yields. Another example of the use of predicted responses to index selection in extension is to alleviate concerns regarding indexes with a negative weight on SCC as indicators of susceptibility to mastitis. Such indexes are often perceived to reduce SCC, which raises concerns about reaching levels of SCC that are too low to manifest an effective resistance to infection. Consideration of responses to selection on an index that includes production and SCC illustrates that such an index does not result in dramatic reductions in SCC but may instead reduce the rise in SCC that is the result of the positive genetic correlation between SCC and production.

6-40